Merge branch 'develop' into dfeng/fwd_f8_fix

Jenkinsfile

@@ -139,7 +139,7 @@ def buildDocker(install_prefix){
         else{
             echo "Checking for image: ${image_name}"
             sh "docker manifest inspect --insecure ${image_name}"
-            echo "Image: ${image_name} found!! Skipping building image"
+            echo "Image: ${image_name} found! Skipping building image"
         }
     }
     catch(Exception ex){
@@ -213,8 +213,10 @@ def cmake_build(Map conf=[:]){
     if (setup_args.contains("gfx94")){
         invocation_tag="gfx94"
     }
+    echo "invocation tag: ${invocation_tag}"
+    def redis_pre_setup_cmd = pre_setup_cmd
     if(check_host() && params.USE_SCCACHE && "${env.CK_SCCACHE}" != "null" && "${invocation_tag}" != "") {
-        pre_setup_cmd = pre_setup_cmd + """
+        redis_pre_setup_cmd = pre_setup_cmd + """
             #!/bin/bash
             export ROCM_PATH=/opt/rocm
             export SCCACHE_ENABLED=true
@@ -228,18 +230,30 @@ def cmake_build(Map conf=[:]){
             export SCCACHE_C_CUSTOM_CACHE_BUSTER="${invocation_tag}"
             echo \$SCCACHE_C_CUSTOM_CACHE_BUSTER
             stunnel ../script/redis-cli.conf
-            (
-                set -e
-                ../script/sccache_wrapper.sh --enforce_redis
-            )
-            error_code=\$?
-            if [ \$error_code -ne 0 ]; then
-                echo "could not connect to the redis server. using sccache locally."
-                ../script/sccache_wrapper.sh
-            fi
+            ../script/sccache_wrapper.sh --enforce_redis
         """
-        setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER=sccache -DCMAKE_C_COMPILER_LAUNCHER=sccache " + setup_args
+        try {
+            def cmd1 = conf.get("cmd1", """
+                    ${redis_pre_setup_cmd}
+                """)
+            sh cmd1
+            setup_args = " -DCMAKE_CXX_COMPILER_LAUNCHER=sccache -DCMAKE_C_COMPILER_LAUNCHER=sccache " + setup_args
+        }
+        catch(Exception err){
+            echo "could not connect to redis server: ${err.getMessage()}. will not use sccache."
+            def cmd2 = conf.get("cmd2", """
+                    ${pre_setup_cmd}
+                """)
+            sh cmd2
+        }
+    }
+    else{
+        def cmd3 = conf.get("cmd3",  """
+                ${pre_setup_cmd}
+            """)
+        sh cmd3
     }
+
     def setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args}   .. ")
     // reduce parallelism when compiling, clang uses too much memory
     def nt = nthreads()
@@ -247,14 +261,16 @@ def cmake_build(Map conf=[:]){
     def execute_cmd = conf.get("execute_cmd", "")
 
     def cmd = conf.get("cmd", """
-            ${pre_setup_cmd}
             ${setup_cmd}
             ${build_cmd}
             ${execute_cmd}
         """)
 
     echo cmd
-    sh cmd
+
+    dir("build"){
+        sh cmd
+    }
 
     // Only archive from master or develop
     if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) {

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp

@@ -184,7 +184,8 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
                 return false;
             }
         }
-        else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940")
+        else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940" ||
+                ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
                            is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt

@@ -108,13 +108,15 @@ if (ENABLE_PIPELINE_V2_OPT)
        CK_EXPERIMENTAL_PIPELINE_V2_IGLP_OPT=1
     )
 
+    # TODO: The "-vectorize-slp=false" LLVM option is a workaround to prevent inefficient instruction scheduling
+    #       caused by the SLP Vectorizer. Remove this option after fix the SLP Vectorizer issue.
     # layout=NT
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
-       COMPILE_OPTIONS ";;"
+       COMPILE_OPTIONS ";-mllvm;-vectorize-slp=false"
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
     # layout=NN
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES
-       COMPILE_OPTIONS ";;"
+       COMPILE_OPTIONS ";-mllvm;-vectorize-slp=false"
        COMPILE_DEFINITIONS "${WAVES_PER_EU_DEFS};${IGLP_OPT_DEFS}")
     # layout=TT
     set_source_files_properties(device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp PROPERTIES